Objectives

Objectives rank products that already satisfy constraints. A similarity objective names its reference, metric, and direction:

{
  "type": "similarity",
  "reference_id": "aspirin",
  "metric_id": "rdkit.ecfp4_tanimoto",
  "direction": "maximize"
}

Initial similarity IDs are rdkit.ecfp4_tanimoto, cheese.shape, and cheese.electrostatic. CHEESE objectives pin their model release and retain raw component scores. Scientifically material choices are never inferred: shape versus electrostatic, logP versus logD, scaffold method, or an ADMET model must be explicit.

Several references that jointly define one objective remain one selection. Independent result sets for several inputs are a selection_batch run.

Predicted-property shortlist acquisition

A ranked similarity selection can optionally reduce its shortlist using one advertised predicted-property endpoint:

{
  "execution": {
    "quality": "balanced",
    "acquisition": {
      "type": "predicted_property_rank_quantile",
      "endpoint_id": "openadmet-herg-pchembl",
      "direction": "minimize",
      "keep_fraction": 0.25
    }
  }
}

This operation preserves the similarity-led candidate generation and keeps the requested rank quantile in the selected direction. It is intentionally not an objective or hard eligibility constraint: CP16 values are predicted, experimental-acquisition-only signals and do not establish that an assembled molecule meets an ADMET threshold. The API returns per-hit predicted values and applicability plus response-level model and candidate-count provenance.

Only ranked_similarity is supported. Sampled selections, ranked selections without a similarity objective, missing endpoints, and databases whose loaded release lacks the signed CP16 assets fail validation before a credit is reserved.